How to transcribe an interview without losing a day to it

The quality of an interview transcript is set at the recording stage, not the transcription stage. No software recovers a word that was never captured cleanly.

Put the microphone between you, closer to them than to you. You know what you said. The recording exists for what they said.

Record the consent exchange, and record it first. "Are you happy for this to be recorded?" followed by their answer belongs in the file, not in your memory of the file.

Ask them to say their name and role at the top. This gives you a clean voice sample, correct spelling in your own notes afterwards, and a reference for every ambiguous attribution.

Write down the names, companies and technical terms you expect to come up. You will use this list in the checking pass, and having it before the interview means you can also ask them to spell anything unusual at the time.

Resist talking over them. Every overlap is a place the transcript will go wrong, and an interviewer's encouraging noises cost more in post than they gain in rapport.

Upload the whole recording in one piece, timestamps on, and let it run. Splitting a long interview into chunks introduces errors at every boundary and restarts speaker labelling each time.

Do not start correcting while it generates. The instinct to fix the first paragraph you see leads to a document that is polished at the top and unchecked at the bottom, which is worse than one that is uniformly rough, because you cannot tell which is which.

If the interview is sensitive, check the service's retention policy before uploading rather than after. That is a decision about your participant's data, not a technical setting.

Play the audio at 1.5x and read along. This is the whole technique. You are not typing, you are watching for the places the text and the audio diverge, and at that speed a one-hour interview takes about forty minutes to check.

Work through your names list first, before the read-along. Search for each name, fix every instance at once, and you remove the most common error class in a couple of minutes.

Then numbers. Then the short turns, under about three seconds, where the two of you overlapped.

Mark anything you cannot make out as [inaudible] with its timestamp. If you are going to quote near that point, the timestamp tells you exactly where to listen again.

Remove fillers and stammers, change nothing else. Clean verbatim is the standard for interviews because it reads properly without putting words in anyone's mouth.

For one hour of clear two-person audio: a few minutes to generate, five minutes on the names list, forty minutes reading along at 1.5x, ten minutes tidying speaker labels and paragraph breaks. Call it an hour.

Add half again for a strong accent the model has not heard much of, a noisy room, or a subject full of specialist vocabulary.

Double it for three or more speakers, because attribution rather than words becomes the work.

Typing the same hour from scratch is four to six hours. That is the whole argument for the hybrid approach, and it is why nobody should be transcribing from a blank page in 2026.

Put a header on it: date, duration, who was interviewed, who interviewed them, verbatim level, who checked it. None of that will be in anyone's memory later, and all of it decides whether the transcript can be relied on.

Keep the audio next to the transcript with the same file name. A transcript whose audio has gone missing is an assertion, not a record.

If the transcript is for research, do the anonymising pass now, while you still remember which details identify whom.